Medicare Will Pay for AI Heart Scans That Haven't Proven They Work
Medicare just announced they’ll reimburse for AI-enhanced cardiac CT interpretation, which means CMS will reimburse companies like HeartFlow, Euclid, and Cleerly about 5 times more than they reimburse for the actual scan.
The vision these cutting edge cardiac imaging companies are selling is really exciting - I fully understand why there’s so much excitement (and hype) behind these tools.
Instead of just seeing whether you have plaque in your arteries, these advanced scans can theoretically identify “vulnerable plaques” - the inflamed, high-risk lesions supposedly ready to rupture and cause heart attacks.1
The promise is personalized medicine and precision prevention. The holy grail of cardiology.
Here’s what we actually know: These tools predict outcomes in observational studies.
That’s it. We have no idea whether (or how) using them might prevent heart attacks.
Contemporary cardiac scans are useful
Current cardiac CTs are pretty good as is.
They scan the arteries and let us see both calcified and noncalcified plaque, estimate the degree of obstruction of the arteries, and identify high risk findings like spotty calcifications, a napkin ring sign, positive remodeling, and low attenuation plaque.
That’s pretty useful information!
These tests aren’t perfect, but we have clear data that they’re at least as good as stress tests at helping to diagnose and treat cardiovascular disease (and might even be a bit better).
The advantage of a CTA2 over a stress test is that you can define someone’s anatomy and diagnose cardiovascular disease that hasn’t yet led to a severe blockage of an artery.
A stress test will only be positive when an artery is more than 70% blocked, but a CTA can see plaque that’s way less obstructive than that.
That’s part of why there are some cardiologists who consider themselves “CT First” when it comes to evaluating a patient with suspected heart disease.
And so like many areas in medicine where things work pretty well, it shouldn’t be surprising that there are a lot of smart people working on a lot of innovation in the cardiac CT space.
The next generation of CT scans have great promise
These advanced cardiac CT scans can utilize information about plaque morphology, plaque composition, coronary inflammation, and fluid dynamics to help identify higher risk patients.
Eric Topol has a great graphic on his excellent Substack describing the different offerings in the marketplace:
These scans sound great in theory - I mean, who wouldn’t want better data with more precise risk stratification and AI-enhanced imaging?
But here’s the problem: Prediction is not the same as intervention.
These companies have impressive-looking data showing their AI can identify high-risk features. What they don’t have is evidence that using this information to change treatment actually prevents heart attacks or saves lives.
And, perhaps more importantly, they don’t have evidence that this additional data adds anything to patient outcomes beyond what we get from identifying high risk plaque features on a standard CTA.
Do these scans mean we’re just going to send more patients for stents they probably don’t need?
Remember the PREVENT trial? Researchers randomized patients with high-risk but non-obstructive plaques to preventive stenting versus medical therapy. The result? Stenting reduced future catheterizations but didn’t change lifespan or symptoms.3
When I wrote about that study last year, my conclusion was basically “when something seems too good to be true, it probably is.” Check it out here:
Preventative stenting for high risk patients
One of the hardest parts about heart disease is that you can’t predict who is going to have a heart attack with much precision.
That’s not exactly a ringing endorsement for intervening on vulnerable plaques.
After all, about 90% of patients in that study didn’t need a stent for the entire duration of follow up.
So it’s easy for me to envision a world where these advanced scans are easy to get, well reimbursed, and only ends up being used to send more patients to the cath lab rather than treat them with the types of medical therapies that will actually reduce heart attacks.
I have real concern that we’re just going to use these tools to just put a whole bunch of stents in patients who didn’t actually need them.
What’s Actually Missing
We aren’t perfect at evaluating cardiac risk, but we’re actually pretty good at it.
My perception is that the unmet need isn’t really better diagnostics, it’s better risk factor control and more thoroughly applied medical interventions.
More than 99% of people who have cardiac events had uncontrolled standard risk factors before:
If you add in an imaging test like a standard CTA, you’re not going to miss many people who are at risk.
The optimistic theory is that these scans help us better discriminate which patients with uncontrolled risk factors need aggressive treatment and which don’t. But there’s a leap between hypothesizing these tools can do that and proving they actually prevent heart attacks or save lives.
Because the gap I see in practice isn’t that we can’t figure out who to treat, the gap is that it’s hard to optimally control everything.
Seeing a high risk CT scan doesn’t mean that we are going to more effectively persuade people to take blood pressure meds or statins.
The most important question with any diagnostic test is what do you do with the information
The critical question isn’t whether AI can find more high-risk features on a CT scan. It’s whether finding those features changes what I do in a way that improves outcomes patients actually care about.
Does identifying a lipid rich necrotic core mean you should get a PCSK9 inhibitor instead of just a statin?
Does peri-coronary inflammation mean you should you start an anti-inflammatory therapy?
What total plaque volume or morphology means you would benefit from aspirin for primary prevention?
We have no idea about the answers to any of these questions.
Shouldn’t these tools have to prove that they make outcomes better or give our management decisions more clarity before we all pay for them?
And until we know how helpful they are, we’re creating a pathway for something I’ve written about before: asymptomatic patients getting scans, getting labeled as high-risk based on plaque features, and ending up with unnecessary stents.4
This is totally absurd from a financial perspective
How come none of these companies need to show that they make outcomes better before we start subsidizing these companies with our tax dollars?
Because once Medicare starts paying for it, private insurance is likely to follow suit, which raises our insurance premiums.
I’m not saying these tools are dangerous.
I’m not saying they shouldn’t be FDA approved - they clearly work as advertised in identifying imaging features.
What I’m saying is: shouldn’t we know if they actually help before we all start paying for them?
The bar for medical technology that we all agree to pay for should be higher than “this correlates with outcomes in retrospective studies.”
It should be “this changes outcomes when we use it to guide treatment.”
We should be able to use these tools, but we should also study them (and everyone shouldn’t automatically have them covered)
I have real concern that we are going to end up using these scans to overreact to concerning findings that should be managed with medications and treat patients with stents instead of statins.
These companies should be doing the randomized trials.
A few suggestions:
IMPACT (Informed Management with Plaque Analysis in Cardiac CT). Does knowing actually change anything? Randomize patients already getting cardiac CTs to standard interpretation versus AI-enhanced vulnerable plaque analysis. Both groups get guideline-based care. Track whether doctors treat people differently when they see the fancy plaque features, and whether those different treatments prevent heart attacks. This tests the real question: does the extra information from the AI enhancements actually help beyond what we currently get?
TAILORED (Targeted AI-guided Lipid and Inflammatory Optimization using Risk-Enhanced Detection. Does imaging beat clinical risk scores? Screen high-risk primary prevention patients (no prior heart attack or stroke). Randomize to clinical risk score guided treatment versus CT + AI interpretation guided treatment. If imaging-guided personalization beats our current approach of treating everyone at intermediate risk the same way, these scans might be worth it. If not, we’re just replacing one imperfect system with a more expensive imperfect system.
PREVENT 2.0 - should we stent the most inflamed, highest risk plaques on advanced CT? PREVENT showed stenting non-obstructive lesions doesn’t help. But maybe if you’re more selective - only stenting plaques with active inflammation or super high risk features - you’d see benefit? Randomize patients with these specific vulnerable features to optimal medical therapy versus OMT plus preventive stenting. A neutral trial here might actually make us stop barking up the “vulnerable plaque” tree.
That would be hard. And expensive. And take years.
But that’s what we should demand before turning on the reimbursement spigot.
Instead, we’re putting the cart before the horse - again.
We’re about to flood cardiology practices with advanced CT interpretation, leading to more testing, more “findings,” more interventions, and quite possibly no improvement in the outcomes that matter.
The future of cardiac CT imaging might be transformative. It might help us personalize prevention and identify who truly needs aggressive therapy. But right now, it’s a solution in search of evidence.
And so we might be living in a world where your tax dollars are going to pay for cardiac CT analyses that might not do anything other than line the pockets of the people who invented them.
What do you think? Are we being too cautious about promising technology, or are we repeating the same pattern of paying for tests before we know if they help? You can always hit reply if you don’t want to comment - I really do read every response.
The reason “vulnerable plaques” is in quotes is because some people think that it’s a myth. See this article, aptly titled “The Myth of the Vulnerable Plaque.” That paper is a 10 year old look at the concept that a focus on the specifics of any one individual plaque (either the way it looks or how tight it is) may be the wrong way to think about heart disease. The argument is that rather than focus on the plaque, we focus on what the plaque means, which is evidence that a patient is high risk overall. The reason the distinction between plaque and patient matters gets to the heart of this newsletter - when you look at the plaque, you may end up employing local strategies, like angioplasty or stents, but when you look at the patient, you employ systemic strategies, like medications that lower lipids or treat high blood pressure or inflammation.
This is different than a calcium score, which only visualizes plaque that has calcium in it rather than visualizes all the plaque that’s there. That gap - a calcium score can’t see soft plaque - is part of why a score of 0 isn’t a particularly useful piece of information, particularly in a young patient.
To be fair, reducing future catheterizations isn’t nothing. But it’s also not the outcome we should be optimizing for when we’re deciding whether to put metal in someone’s arteries.
Just a quick aside on the bear case for these scans. While I don’t think that these scans are very likely to be bad for patients, I also don’t think it’s insane to think about the potential downside. The worst case with widely implementing these scans into our predictive pathways is that they may falsely reassure people with “low risk CTs” that they don’t need to be on medications that would reduce their cardiovascular risk. It’s certainly true that these scans seem to identify the highest risk patients, but unfortunately the real life biology is almost always more complicated than our models suggest that it is. Is it that implausible to think that it’s possible to have a heart attack even if your plaque appears low risk with Cleerly’s AI model?
It is mind-blowing (and probably involves corruption somewhere) that medicare would pay for this in the absence of outcome data.
Surrogate outcomes (like "vulnerable plaque" on a CT) MIGHT correlate to actual clinical outcomes that anyone actually cares about….but the burden is on proponents to actually prove that.
This is what makes me freaking livid. We routinely pay for medicines for dementia that "do not not work" ( my way of pointing out how low the bar is in approving dementia meds) and also pay for medicines for rare diseases that essentially don't work as well,but I can not get patients on GLPs or SGL 2s, which are game changers. Honestly it makes me go to a dark place sometimes. Again just another example of medical waste.